Using Transformers to Forecast Incredibly Rare Solar Flares

Forecasting incredibly rare events like major solar flares presents unique machine learning challenges that require a shift in modeling strategy. The process involves using magnetic field data from NASA's Solar Dynamics Observatory and performing significant feature engineering to capture relevant magnetic structures. Due to the extreme class imbalance, standard metrics like accuracy are insufficient, necessitating the use of the True Skill Statistic (TSS) and the creation of a "tail model" that exclusively trains on rare event data. This specialized tail model is then combined with a general distribution model using a Transformer architecture to produce a more robust and accurate forecast for these catastrophic events.